Thomson Reuters Didn't Build Its AI. It Bought the Startups and Plugged Them Into a Moat.

In the summer of 2023, Thomson Reuters wrote a $650 million check for a company with 104 employees.² The startup, Casetext, had raised about $64 million across its entire life as a venture-backed business.⁸ So Thomson Reuters paid roughly ten times the capital the company had ever raised — in cash — for a flagship product, CoCounsel, that it could have spent that money trying to build itself. It chose not to. That choice is the whole strategy, and almost nobody read it correctly.

The story in the headlines was a transformation: a centuries-old information company reinventing itself for the age of generative AI. The real story is quieter and far more deliberate. Thomson Reuters did not reinvent itself. It went shopping — and it shopped in the only aisle where its existing assets made the purchases unfair to everyone else.

The thesis is this: Thomson Reuters' AI pivot is not a transformation at all. It's a balance-sheet-backed adjacency roll-up, where the acquired AI is the cheap part and the moat is the proprietary content it gets bolted onto.

The phrase that gives away the playbook

On June 26, 2023, in the SEC filing announcing the Casetext deal, CEO Steve Hasker described the company's approach in three words: 'build, partner and buy.'¹ The press would later attach that phrase to an investor day, as though it were a freshly minted vision. It wasn't. It was already on the record the day the first acquisition was announced — and the order of the words matters. Build comes first rhetorically, but buy is what actually happened. Casetext was the buy that proved the playbook.

Here is the move worth seeing clearly. Casetext had something Thomson Reuters wanted but did not have: early access to OpenAI's GPT-4, and a working AI legal assistant built on top of it.⁸ Thomson Reuters had something Casetext could never build: Westlaw, Checkpoint, and decades of proprietary legal and tax content that competitors cannot simply scrape together. The acquisition fused the new layer onto the old corpus. TechCrunch reported that Thomson Reuters framed the deal partly as a way to keep the technology away from competitors — LexisNexis above all.⁸ Buying Casetext wasn't only addition. It was denial.

Why the AI is the cheap part

Strip away the narrative and look at what a generative-AI model actually is in this market: a layer that turns a question into an answer. Anyone with capital and engineers can rent the model — Casetext rented GPT-4. What no one can rent is the thing the model has to read to be useful to a lawyer or a tax preparer: the authoritative, structured, litigation-tested content corpus that Thomson Reuters has spent generations assembling. An AI legal assistant trained on the open web hallucinates case law. An AI legal assistant pointed at Westlaw answers like an associate. The model is interchangeable. The corpus is not.

This is why the install base is the asset, not the algorithm. Casetext arrived with more than 10,000 law firm and corporate legal department customers,² but those relationships are dwarfed by the ones Thomson Reuters already owned. The acquisitions don't open a new market — they drop a new product into a market the company already dominates, with billing relationships already in place. That's what makes it an adjacency expansion rather than a pivot. A pivot abandons the old ground. This defends it.

The same trick, run again in tax

If Casetext were a one-off, you could call it a bet. It wasn't. On January 2, 2025, Thomson Reuters paid $600 million in cash for SafeSend, a tool that automates the 'last-mile' of the tax return — and SafeSend was already used by 70% of the top 500 U.S. accounting firms.⁴ Read that penetration figure twice. Thomson Reuters did not buy a promising startup hoping to win the accounting profession. It bought a company that had already won a large share of it, then plugged it into Checkpoint, its existing tax-and-accounting franchise. SafeSend was projected to generate roughly $60 million in 2025 revenue and grow in excess of 25% annually.⁴ You are buying installed distribution, not potential.

And in the fourth quarter of 2024, the company added Materia, described as a specialist in agentic AI for the tax, audit and accounting profession.⁶ The price was never disclosed in any filing — a detail worth flagging, because the absence is real, not coyness covering a fabricated number.⁶ Three acquisitions, two professions, one pattern: take a near-monopoly install base, find the AI capability you lack, and buy it before a rival can.

Isn't this just a big company outspending its rivals?

The fair objection is that anyone with enough cash can buy startups, so this looks less like strategy and more like a checkbook. There's something to that. Thomson Reuters had the means: it sold its remaining LSEG stake in May 2024 and divested FindLaw that December,³ freeing capital to spend. But cash alone doesn't explain why the deals work. A balance-sheet-rich company with no relevant corpus would be buying AI it couldn't distribute. The reason these acquisitions compound is that each one lands on ground Thomson Reuters already owns — the content, the customers, the billing relationships. The check is necessary; the install base is what makes the check pay off.

The honest counter is harder. AI models are getting cheaper and better fast, and at some point a rival could point a strong-enough model at good-enough public data and approximate the answers without the proprietary corpus. That's the real threat to the thesis, and it's why the spending has to keep going — more than $200 million in 2024, expected to continue at that pace in 2025.⁵ But the early results suggest the moat is holding so far. Thomson Reuters grew organic revenue 7% in 2024, ahead of its own guidance,³ and in Q3 2025 organic revenue grew 7% again while adjusted EBITDA margin expanded to 37.7%, on net leverage of just 0.6x.⁷ A company being commoditized does not expand margins while it spends.⁷

Thomson Reuters didn't reinvent itself in the age of AI. It did something less dramatic and more durable: it recognized that the model was the cheap, rentable, copyable part — and that the only thing nobody could buy was the corpus it had already spent generations building. So it let other people build the assistants, then bought them and pointed them at the one library no competitor could assemble. The genius was never the AI. It was knowing which part was the moat, and refusing to confuse it with the part everyone else was racing to build.